An Optimal Feature Extraction using Deep Learning ...

Received: July 21, 2020. Revised: August 26, 2020. 318

International Journal of Intelligent Engineering and Systems, Vol.13, No.6, 2020 DOI: 10.22266/ijies2020.1231.28

An Optimal Feature Extraction using Deep Learning Technique for Trojan

Detection and Validation using Game Theory

Priyatharishini Murugesan1* Nirmala Devi Manickam1

1Department of Electronics and Communication Engineering, Amrita School of Engineering,

Coimbatore, Amrita Vishwa Vidyapeetham, India * Corresponding author’s Email: [email protected]

Abstract: In modern electronic design, hardware Trojan has emerged as a major threat in the hardware security. To

detect the hardware Trojan is a major problem in testing process because of their inherent concealed nature. In this

work, we propose a deep learning-based Trojan classification approach, which extracts the optimal feature to indicate

the nets affected by the Trojan module. In this approach, a handcrafted algorithm along with the structural report is

also analyzed for extracting further features of the gate level netlist, which stamp out the requirement of golden chip.

This detection technique is also validated using game theoretical approach, which is modelled as zero-sum game

between the attacker and the defender. The Simulation is employed on ISCAS’85, ISCAS’89 and Trust-HUB circuits

and the deep learning algorithm performs the best in detection and classification of Trojan type with an average True

positive rate of 96.69% and an accuracy of 96.25%.

Keywords: Hardware trojan, Hardware security, Deep learning, Feature, Game model.

1. Introduction

Hardware designers need to depend on third party

manufactures due to the technological growth of the

Integrated circuit(IC). As a result, it makes easy for

the attacker to intrude the design at various points.

Such a violation in the security of the IC design may

result in malicious modifications referred to as

Hardware Trojans (HT) [1, 2]. In order to evade from

the formal design testing and verification phase, the

hardware Trojans are modelled in such a manner that

they are stealthy in nature and triggered by rare

events.

Detection of hardware Trojans and prevention of

Trojan insertion by the adversary is a challenging task

and addressed in the recent literatures. Various

detection schemes are broadly classified into side

channel analysis, online checker, functional test,

runtime monitoring. In [3] a vulnerability analysis of

the digital circuit is performed at the behavioural

level circuit. The dummy scan flip flops are intruded

in the design to improve the triggering frequency of

the transition probability. The region-based approach

is one of the classifications of side channel analysis

[4, 5], in which the circuits under test are partitioned

into blocks for detecting the Trojan. In side channel

analysis, the frequency and power profile are

considered, in which the fluctuation in these

parameters for Trojan nets will be very low and is

undetected during testing phase. The solution for the

above drawback advances the scheme of online

checker and here the Trojan prevention schemes

obfuscate the original design from the attackers by

inserting extra module to the design [6]. This process

increases the trustworthiness, but the design and

resource overhead are considered high.

Therefore, in order to address both the drawbacks,

run time detection techniques is emerged with the

concept of reversible logic incorporated in the design

[7]. An optimal test pattern generation technique is

presented in [8], which selects the sparse test set for

detecting the HT. Hence to solve the drawback of

above, the deep learning approach is emerged for

extracting the optimal set of internal net features in a

suitable manner to trigger the Trojan module and this

approach is independent of test patterns.



Figure. 1 Application scenario of deep learning technique

Deep learning algorithm is proven to be the best

method of feature extraction for various real time

applications in image processing, traffic

identification and bio imaging. The Hardware

Trojans can be embedded from register transfer level

to designing gate level netlist, to packing the

integrated chips are shown in the Fig. 1. In the design

phase, the EDA tool is used by the IP suppliers and

chip designers but the reliability of that tool is un-

trusted. In the production phase, the un-trusted

employee’s involvement in design layout needs more

attention. The threat models are inserted to change

the functionality of the circuit by inserting, deleting

or modifying the Trojan components to extract and

control the original chip.

This paper presents an efficient feature extraction

technique for detecting the hardware Trojan in the

gate level net-list. This technique is referred to deep

learning based hardware Trojan detection. Further it

is validated using game theory. The proposed method

comprises three different phases: Optimal Trojan

feature extraction phase which extracts and encodes

the Trojan features. Detection phase consists of

learning phase to train the model and classification

phase to precisely cluster the data, Validation phase

to investigate the proposed detection scheme. The

proposed methodology is evaluated on the

benchmark circuits considering the performance

metrics resembling True positive rate, True negative

rate, Processing time, Probability of Defender,

Expected pay off of defender, Attackers probability

and Accuracy.

The main contributions of this work are as follows:

1) The proposed algorithm is attempted to

extract the efficient features of Trojan nets

from the large pool of nets in the circuit. This

scheme extracts compact list of features,

resulting minimal processing time for

computation.

2) An algorithm is developed to suit any type of

circuit by extracting the handcrafted features.

A specific threat model is considered and the

features are also extracted from synthesis

tool.

3) An automated algorithm is developed along

with K-means clustering to have a great

impact on deciding the normal nets and the

Trojan nets. On applying the proposed

scheme to Trust- HUB circuits a high true

positive rate is achieved and the results are

compared.

4) A net scoring algorithm is developed to

identify the efficient internal nodes in the net

list and to generate the pay off matrix for

validation.

5) The Trojan detection process is validated

using game theory approach with a defender

and adversary as a two-person strategic game

model, to achieve high reliability.

The rest of the paper is organized as follows. The

related work on different schemes for detecting

hardware Trojan is described in section 2. In Section

3, the proposed deep learning approach for detecting

the Trojan is presented. Section 4 describes the

results and analysis for various ISCAS benchmark

and Trust-HUB circuit. Finally, Section 5 concludes

with some future scope on the related topic discussed

in this paper.

2. Related work

The Trojan circuits are analyzed in various levels

of the design, which include significant node

selection and extraction of specific features of the



nets. Based on the outcome of these parameters,

classification of the circuit under test is done using

different classifiers such as support vector machines,

random forest classifiers etc. The different types of

hardware Trojans and its threat models are analyzed

in [9-11]. In [11], hardware Trojan nets are identified

by the proposed Support vector machine based

classifier. The static detection process is discussed

her, which does not require any test patterns for

activating the Trojan model and also avoids the use

of reference golden chip for the analysis. Low

probability of detection rate when large data sets are

involved is its limitations and it is also immune to

noise signal. The Trojan features from the Trojan nets

are extracted in [12] and a random forest classifier is

applied to obtain the optimal set of Trojan features

from the nets. Higher computational power is its

limitations and requires more resources, since the

model involves a lots of tree structure

Classification of Hardware Trojan using multi-

layer neural network is discussed in [13]. The

processing time is more to train the dataset for

decision making which is its limitations since the

hidden layers are based on feature set. In this paper

[14], modelling an artificial neural network (ANN) is

performed in-order to address the pattern recognition

challenges. The statistical indicators metric is

introduced for evaluating the performance of ANN

models for various methods. The ANN model

requires parallel processing power and is suitable for

the numerical problems. Hence all the problems are

to be converted to numerical values and feed to the

model is the limitation of this method. In [15], a

boundary net structure based Trojan classification,

which used machine learning for initial classification

of segregating the nets into Trojan and normal nets.

The extraction of features is done based on the

classification result are its limitations and the

misjudgements of the nets are identified. An

unsupervised learning approach is proposed in [16],

in which local outlier factor (LOF) combined with

Principle component analysis algorithm PL-HTD is

adapted to visualize the Trojan nets. The theoretical

approach is limited to the feature selection in

different phase and is to be analysed. Hence the

complexity of the system increases as scheme

involves different filtering process. In [17-18], a deep

learning algorithm is developed to overcome the

limitations of machine learning, which extracts the

features by itself and the nets are optimally classified

based on the extracted parameter for large dataset

with minimal computational complexity.

In [19] a security based game model is developed,

which aid to the detection process of hardware Trojan.

But the utility function is required to identify the best

set of strategy for the threat model which is its

limitation. An iterated elimination of dominant

strategy is preferred in which one player dominates

the other player to yield better payoff. The testing of

digital circuit is also performed by modelling the

game theoretical approach in [20-21]. A zero sum

game between the attacker and defender is designed

for testing the Trojans, where the defender fails to

detect Trojan pays a high fine value is its limitation.

A novel game model is framed in [22] to analyze the

relation between the manufacturer (attacker) and the

IC testing unit (defender). In this paper [23], a

strategic game is modelled with two players to

determine the suitable strategy for both defender and

intruder. A software framework is also presented,

which provides the mathematical exploration and

also provides the solution of the game.

3. Proposed methodology

A deep learning based hardware Trojan detection

technique is proposed in this paper and the Fig. 2

illustrates the design flow of the scheme.

For the gate level net list, the algorithms are

developed to generate the optimal features of the

internal nodes that can probably excite the Trojan

sites. Golden circuit net list and the Trojan infected

net list are synthesized using Synopsys. The proposed

deep learning approach computes the optimal internal

net features for enhancing the detecting probability

and it is also validated using game theory approach.

The different phases of the proposed scheme are:

Figure. 2 The proposed methodology design flow



Optimal Trojan feature extraction phase, Detection

phase and Validation phase.

3.1 Optimal trojan features extraction phase

In this phase, a gate level netlist for a specific

net in ISCAS and Trust- HUB benchmark circuits are

chosen for selecting optimal features. The complexity

of the model is minimized by the efficient feature

selection phase, which speeds up the learning phase

of the classification algorithm.

3.1.1. Handcrafted feature extraction

The handcrafted algorithms are developed to

extract the certain critical parameters like primary

input, primary output, transition probability, level,

connectivity and toggle rate.

Primary input (PI), primary output (PO):

The minimum level from the input node to the

target net ‘n’ is PI and the minimum level in the

design from the output node to the target net ‘n’ is PO.

Transition probability (TP):

The pseudo code 1 is developed to compute the

transition probability of all net in the logic network.

The nets with low transition probability are extracted,

in which the malicious modules are more likely to be

inserted.

Pseudo code 1 Determining Transition probability

(TP ) value

1: Scan the net list.

2: Export the PI, PO and the type of gate.

3: Initialize the primary input signal probability as

0.5

4: for each net in the netlist do

5: Determine gate type

6: Apply the probability equation at the output of

each net

7: Compute the overall transition probability

values for each net by

TP = Probability of logic 0 × Probability of logic 1.

8: Store TP value of each net in an array.

9: end for

Level (LE):

The number of logic gates that are cascaded

between primary inputs to the target net’n’. This

feature is more likely to increases during malicious

module insertion.

Connectivity(CO):

Connectivity provides the total number of nodes

to which a target net’n’ is connected. If the target net

triggeres on rare event, then this feature will impact

the node associated with target net which may lead to

functional change or leak of information.

Toggle rate(TO):

The toggle count features is the total number of

transistions of a target net’n’for different range of test

vectors. The net with maximum toggle rate are more

sussecptable nodes and these nodes are more likely

for trojan insertion.

In order to confirm the feature selection, Table 1

summarizes the average values of the above specified

features for the genuine nets and Trojan nets for the

benchmark circuit. The outcome of this phase will be

a set of nets with efficient features. It is inferred that

the average value of the Trojan nets seems to be

larger when compared to genuine net. The nets with

high fan in count are likely to be malicious nets, but

some of the genuine nets also have maximum fan in

count. Similarly, it is also observed that, the primary

output is large for the genuine nets compared to

Trojan nets. Hence only with the hand crafted

features classifying the malicious nets are not

sufficient, so the structural report is also analysed to

extract the features relevant to the Trojan types.

3.1.2. Structural and power analysis:

The structural and power reports are generated

using Synopsys tool to extract the relevant features

for the genuine and malicious infected net-list.

Logic gate fan in (FI) and fan out (FO):

The total number of inputs of the logic gate away

from the target net ’n’. The Trojan modules are

activated only when it satisfies the triggering

condition or on rare occurrence of the internal states.

The total number of outputs of the basic gate away

from the target net ’n’. Nets with maximum fan in and

fan out are more likely to be infected nets but this

itself is not enough to distinguish the Trojan nets

from genuine nets.

Table 1. Average value of the feature

Net

Type LE CO PI PO

Fan

In

Fan

Out

Trojan

nets 4.5 1.34 4.5 3.5 1.25

1.5

Genuine

nets 1.5 1.3 1.5 6.5 1.0

1.3



Table 2. Average value of the features based on synthesis

tool Circuit Net Type / Features

Genuine net

Trojan net

TN 0.29 0.46

SPR 0.48 0.64

TR 0.09 0.17 SP 0.007 0.043

LO 0.29 0.46

R 0.02 0.03

PN 2.1 2.7

Net load (LO) and pins (PN):

A device connected to the signal source consumes

some power, which affects the circuit performance

with respect to the primary output. The malicious

module insertion may also lead to variation in the net

load feature and the pins in the design.

Static probability (SPR):

This feature refers to the target net ’n’ pertaining

to the expected state of logic -0 or logic-1 and the

maximum value of static probability are more prone

for Trojan sites because of the rare occurrence of the

logic state.

Resistance(R) and switching power (SP):

A resistive element connected between the load

and the power source, which is used to monitor the

power performance for varying load conditions.

Table 2 intimates the average value of the original

nets versus Trojan nets for the features obtained as a

result of applying the netlist to the structural analysis

using the synthesis tool. It is observed that, the values

in genuine nets are comparatively lesser than

malicious nodes. The outcome of this phase will be a

set of nets with efficient features which are more

feasible to classify the anomaly in the design and are

referred as Optimal Trojan Feature set (OTF).

3.2 Detection process

The OTF for each target net ‘n’ are extracted, but

it is hard to fix the threshold value for the features to

classify the malicious net from Trojan infected.

Hence a deep learning based malicious module

detection method is proposed, where it can

automatically learn and extract the efficient features

for the malicious and genuine nets from the OTF. It

also classifies an unknown gate level net-list into a

set of Trojan nets and genuine net using multilayer

perceptron classifier. The Pseudo code 2 describes

the flow of the proposed detection and classification

of Trojan, which is composed of learning phase and

classification phase.

Pseudo code 2 Detection and Classification of

Trojans using Deep learning algorithm

1: Read the features list of the circuit.

2: Optimize the training features and set to X-train

3: Obtain the unsupervised feature representation of

the data as input to auto encoder.

4: Generate a neural network that are densely

connected for encoding the features

5: Generate a neural network that are densely

connected for decoding the features

6: Decode the training data for extracting best set

of features

7: Apply k-means algorithm for detecting the

presence of Trojan.

8: Visualize the clusters with its labels

9: Implement Multilayer perception for classifying

multiple Trojan types.

Learning phase:

In learning phase, the deep learning architecture

is learned by many known malicious nets and genuine

net by using the extracted feature values. For each

net’n’ the feature value are extracted and it is

considered in a thirteen dimensional feature

vector ’yn’, which is provided to the input layer of the

deep learning architecture. The back propagation

algorithm is employed in order to approximate the

feature vector by minimizing the error value. The two

level hidden layer is developed in the architecture,

which densely connect the feature vectors of the input

layer to that of the output layer. The feature vectors

are encoded in the hidden layer using auto encoder

and the encoded features are also labelled using k-

means clustering algorithm. The labelled feature data

pool are provided to the multilayer perceptron(MLP)

classifier for training.

Classification phase:

In classification phase, the deep learning

algorithm considers the extracted features as the

training data and classifies the unkown gatelevel

netlist into genuine nets and set of Trojan nets.In

order to normalize the training data an unsupervised

autoencoder algorithm is developed and the the

structure of the auto encoder is created using a

densensly connected encoder with Relu as activation

function. The data from the encoder is provided to

densely connected decoder which uses sigmoid as an

activation fuction to decode it. The clustering of

decoded data into as a genuine net or trojan net based

the extracted features of the unknown net-list is

performed using K-means algorithm. The result of

the classification phase will be clustering the

extracted features into original net or trojan nets. The



set of features is passed on to the multilayer

perceptron only if the classifier identifies the

extracted features as Trojan nets and it clusters the

trojan infected nets into combinational, sequential

and always on Trojan. Thus, the detection happens

between genuine or Trojan and also between the type

of Trojans there by increasing the overall reliability

of the system.

3.3 Game theory based validation

The proposed methodology in the Fig. 2 is

validated using the using the game theoritical

approach for testing the hardware trojan.

Modelling the stratergies

In this game model, a decision making process is

performed on an non- coorporative stratergies with

two players attackers and tester. The attacker is the

one who inserts the hardware trojan module and to

minimize the validation, they insert single trojan at a

time. The certain types of trojan like combinational,

sequential and always on trojan classes are chosen

for analysis. The attackers stratergies considered as

inserting any one of the trojan from the specified

trojan classes. Thus the attackers have three strateries

(Y) that are denoted as TC, TS, TA. Thus the

corresponding values for the attackers stratergy for

different Trojan classes are represented as S1,S2,S3

and are computed from the score algorithm.The

attackers fine value(F) is also decided from score

computaion and a thershold value (∆F) is also

fixed,which is imposed on the attacker when a trojan

is identified. This model is also considered as a

rational approach, where the attacker intrusion of

malicious circuit minimizes the likelihood for

detection and the defender aims for maximizing the

the payoff value.

Generating the payoff matrix

A mixed strategy Nash equilibrium is employed

in-order to get the solution for the game model and

also it allows us to determine the two player’s

frequency of choosing the strategies at equilibrium

condition. A payoff matrix is generated using the

score value of the Trojan inserted circuit along with

the fine values to be imposed on the attackers. For

this illustration, the score value of the attacker’s

strategy is computed from the score algorithm which

is performed by adding all the features for each net.

The maximum value (M), minimum value (N) from

the nets is considered and the score value is clustered

into number of attacker strategy groups by Eq. (1).

𝑆𝑖 = [𝑀 +𝑁

𝑌× (𝑖)] 𝑓𝑜𝑟 𝑖 = 1,2, . . 𝑌 (1)

A sample circuit of ISCAS 85’ C17 bench mark

circuit the maximum net value M=7, minimum net

value is N=4, the score value is computed by equation

1 to obtain the clusters S1 =4, S2=8, S3 =12. In this

model, the defender considers a mixed stratergy in

which two trojan types are tested simultaneously out

of three trojan classes and therefore the defender has

3C2 possible stratergies represented as TCTS, TSTA,

TCTA. The fine value F is also decided for the

corresponding circuit based on the Eq. (2).

𝐹𝑖𝑛𝑒(𝐹) = 𝑚𝑎𝑥(𝑆𝑖) + ∆𝐹 (2)

where ∆F is choosen as 2 inorder to get maximum

positive payoff for the defender. The attackers and

defenders payoff for various fine values are also

analysed. The Table 3 specifies the pay off matix

generated for the sample ciruits of C17 and C6288 of

ISCAS’85 benchmark circuit.

Determining the expected pay off

The game model is incorporated using mixed

strategy Nash equilibrium and the best response of

the defender to the attacker intruding a sequential

Trojan of type TS is to play TCTS or TSTA. So that

the malicious module can be identified and the

attacker can be imposed with a fine ’F’ to get a

negative pay off value to the attacker and positive pay

off value to the defender. If the defender choice is to

play TSTA then the attacker’s best choice is to play

TC and go undetected. Thus the pure stratery nash

equilibrium is not suitable for the game model shown

in Table 3 because of its circular reasoning. The

defender chooses the mixed stratergy to get the

maximum payoff against the attacker, who seeks to

intrude trojan that is undetectable with maximum

impact. From the pay off matrix, the probability is

computed by the online mixed stratergy solver for the

defender and attacker. The probability of defender’s

strategies TCTS, TSTA, TCTA are represented as p1,

p2, 1-p1-p2 and that of attackers strategies TC, TS,

TA are represented as q1, q2, 1-q1-q2. The expected

payoff values for the player’s are calculated using the

Eq. (3).

Table 3. Pay off matrix generation for hardware trojan

detection

Bench mark Circuit

Defender (C17) Defender (C6288)

TC TS TS TA TA TC TC TS TS TA TA TC

Att

ack

er TC -F,F 12,-12 -F,F -F,F 63,-63 -F,F

TS -F,F -F,F 8,-8 -F,F -F,F 42-42

TA 4,-4 -F,F -F,F 21,-21 -F,F -F,F



𝐸𝑃 = (𝑆1 × 𝑝1 × 𝑞1) + (𝑆2 × 𝑝2 × 𝑞2) + (𝑆3 × (1 − 𝑝1 − 𝑝2) × (1 − 𝑞1 − 𝑞2)) (3)

4. Result and analysis

The proposed methodology is validated using

ISCAS’85, ISCAS’89 and Trust-HUB circuits. An

optimal feature set is generated using deep learning

algorithm for the circuit under test and the nets with

varying feature values are likely to be Trojan

triggered nets. The threat models intruded into the

benchmark circuit are combinational (Tc), sequential

(Ts) and always on Trojan (TA) modules. These

threat modules alter the circuit functionality when it

is triggered on rare conditions. Synopsys DC

compiler is used to synthesis the infected design and

the golden design.

Table 4 illustrates the comparison of different

features values of the genuine nets and Trojan

infected nets, obtained as a result of intruding a ring

oscillator based Trojan module in the design. The

outcome of extraction phase will be selecting a set of

Trojan nets with varying feature value for the

benchmark circuits. It is observed that when the

Trojan module is intruded in the design, the feature

values of the corresponding Trojan nets are varying

with respect to the normal nets. Thus a sample of C17

and C432 circuit is shown along with its Trojan nets

represented as NT, which shows the optimal features

are extracted for detecting the Trojan module in the

design for high reliability.

The maximum values of the extracted features are

listed in Table 5 and the analysis is done by inserting

the three different types of Trojan individually in the

benchmark circuit. It is observed that the maximum

values of the extracted feature of the Trojan circuits

TC, TS, TA is comparatively higher than the genuine

circuits. It is observed that the variation in extracted

optimal feature values indicate the presence of Trojan

module in the design.

Table 6 reveals the processing time for executing

the deep learning algorithm, achieved as a result of

inserting (i) Tc, Ts, TA Trojan modules individually

and (ii) all the Trojans at the same time for the

benchmark circuits. It is observed that for the C7552

circuit which has large number of primary input and

nets compared to other circuits consume only an

average processing time of 9.02 s and for S13207 it

takes 13.56 s. Thus the Table 6 depicts that the

average processing time required for computing the

deep learning algorithm for complex circuits is less.

It is also inferred that the extracted optimal features

require less processing time, with minimal

computational complexity for detecting the hardware

Trojan.

The Fig. 3 shows the total number of nets affected

by the insertion of Trojan module for ISCAS’85,

ISCAS’89 and Trust-HUB circuits. The

combinational (Tc), sequential (Ts) and always on

Trojan (TA) are designed in order to observe the

feature values for specific trigger conditions

Table 4. Comparison of different features values of the genuine nets and trojan infected nets

Benchmark

circuit

Nets LE CO PI PO FI FO LO RE PN TN SPR TO SPO

C17

N1 0 1 0 3 1 1 0.24 0.02 2 0.243 0.247 0.0974 0.0058

NT1 0 1 0 8 1 1 0.24 0.02 2 0.243 0.616 0.123 0.0073

N19 2 1 2 1 1 1 0.24 0.02 2 0.243 0.616 0.123 0.0073

NT19 7 2 7 1 2 1 0.54 0.04 3 0.544 0.616 0.123 0.0164

N23 3 1 3 0 1 1 0.24 0.02 2 0.243 0.568 0.1447 0.0086

NT23 8 1 8 0 1 1 0.24 0.02 2 0.243 1 0 0

C432

N118 1 1 1 16 1 1 0.24 0.02 2 0.243 0.284 0.1881 0.0112

NT118 6 2 6 21 2 1 0.24 0.02 2 0.243 0.5 0.1 0.006

N119 1 2 1 16 2 1 0.54 0.04 3 0.243 0.5 0.1 0.006

NT119 6 4 6 21 4 1 0.54 0.04 3 0.544 0.5 0.1 0.0133

N123 1 2 1 16 2 1 0.54 0.04 3 0.243 0.5 0.1 0.006

NT123 6 2 6 21 2 1 0.54 0.04 3 0.544 0.5 0.1 0.0133

N163 2 2 2 15 2 1 0.24 0.02 2 0.243 0 0 0

NT163 7 2 7 20 2 1 0.24 0.02 2 0.544 0.76 0.0949 0.0126



Table 5. Maximum values of the extracted feature of the benchmark circuit

Bench mark

circuit

LE CO PI PO FI FO LO RE PN TN SPR TO SPO

C17 3 2 3 3 2 1 0.54 0.04 3 0.544 0.753 0.1447 0.0164

TC 3 2 3 3 3 1 0.86 0.07 4 0.544 0.749 0.1921 0.0159

TS 5 4 5 5 4 1 1.2 0.09 5 1.203 0.757 0.1437 0.0358

TA 8 2 8 8 2 2 0.54 0.04 3 0.544 1 0.1432 0.0164

C432 17 9 17 17 19 1 9.88 0.76 20 9.876 0.921 0.2784 0.493

TC 18 9 18 18 20 1 9.88 0.76 22 9.876 1 0.2801 0.493

TS 18 9 18 19 21 1 9.88 0.76 22 9.876 1 0.2737 0.4869

TA 17 9 17 17 19 2 9.88 0.76 20 9.876 1 0.2804 0.493

C1908 35 25 35 35 25 1 15.22 1.18 26 15.218 1 0.7531 1.3853

TC 35 16 35 37 26 1 15.22 1.18 27 15.218 1 0.7531 1.3853

TS 37 16 37 40 28 2 15.22 1.18 28 15.218 1 0.7506 1.4025

TA 40 16 40 40 26 1 15.22 1.18 26 15.218 1 1.014 1.3853

C6288 119 34 119 119 16 1 7.48 0.58 17 0.863 1 0.8035 0.1699

TC 120 36 120 120 16 1 7.48 0.58 17 0.863 1 0.8035 0.1699

TS 122 16 122 122 16 2 7.48 0.58 19 1.951 1 0.8077 0.1704

TA 124 16 124 124 16 1 7.48 0.58 18 1.203 1 0.8047 0.1701

Table 6. Processing time for executing deep learning

algorithm

to validate the presence of Trojan. The Fig. 3 (a)

shows the number of affected nets by inserting

different types of Trojan simultaneously to the bench

mark circuit. It is observed that for C7552 circuit, the

number of nets affected by Tc type Trojan are 625,

Ts type are 1122 and TA type Trojan are 450. Thus

the extracted features from the deep learning

algorithm classify the intruded Trojans into its

corresponding Trojan types. Similarly Fig. 3 (b)

shows the number of nets affected by simultaneously

inserting (i) combinational & sequential Trojan, (ii)

always on & sequential Trojans modules in the circuit.

It is observed that the nets are clustered only in its

corresponding Trojan type. The circuit under test is

introduced with only one Trojan type and is shown in

Fig. 3 (c). Thus it is inferred that the optimal features

are extracted, as a results the data sets are classified

to the corresponding Trojan types inserted in the

design which provides high reliability of the system.

The accuracy for evaluating the deep learning

algorithm for the benchmark mark circuit is

manifested in Fig. 4. For the circuit C7552, the

accuracy is 95.4% for Trojan Type Tc, 95% for Type

Ts, 95.25 for Trojan Type TA and has an accuracy of

97.31 % with all types of Trojan inserted. It is

inferred that the accuracy for the different Trojan

types inserted simultaneous is high compared to the

single Trojan insertion. Hence the extracted features

by the deep learning algorithm is capable of

classifying and detecting multiple Trojans in the

design with an average accurate rate of 96.02% for

ISCAS’85 and ISCAS‘89 circuits.

The attackers probability of the game theorical

approach with high score value on combinational

type trojan is shown in Fig. 5. It is observed that the

probability of the attacker inserting the hardware

trojan minimizes the score value of corresponding

Bench mark circuit

No. of inputs

Normal nets

Processing Time(s)

Tc Ts TA Tc Ts TA

C17 5 11 0.4 0.42 0.42 0.48

C432 36 196 1.69 1.71 1.75 1.80

C1908 33 913 1.69 1.71 1.75 1.80

C6288 32 2445 2.17 2.76 2.80 2.83

C7552 207 3720 8.95 9.0 9.03 9.12

S27 4 14 0.1 0.14 0.18 0.23

S298 3 133 0.3 0.32 0.34 0.38

S641 35 433 2.78 2.89 3.10 316

S820 18 309 1.78 1.89 2.10 2.13

S13207 62 8141 13.1 13.2 13.8 14.0



(a)

(b)

(c)

Figure. 3 Number of data points infected by the trojan: (a)

all types of trojan inserted, (b) two types of trojan

inserted, and (c) single trojan inserted

Figure. 4 Accuracy of the deep learning algorithm with

different trojan types

Figure. 5. Mixed strategy Nash equilibrium for attacker

trojan type and high score value trojan modules are

less likely inserted by the attackers as these high

score trojan types are more effectively protected by

the defender. Hence it is observed that the

combinational type which has high score value are

less choosen by the attacker.

The Fig. 6 shows the probability of defender

detecting the hardware trojan in the game theory. The

different trojan types are inserted in the design one at

a time and the probability for the defender is shown

in the Fig. 6 (a). It is observed that the probability of

the defender is high for the corresponding trojan

types, which infers that the game theoritical model

optimally detects the trojan type based on the score

value. Fig. 6 (b) shows the probability value for

Trust-HUB circuits and it is observed that the

defender has maximum probability for the

corresponding trojan type inserted in the trusthub

circuit. The optimal results are obtained by this game

theory model and it proves that the features extracted

from the deep learning algorithm is an optimal set for

detecting the hardware Trojan.

The Expected payoff for the defender is shown in

Fig. 7 in which the fine value assigned to the game

model yields a negative payoff for the attacker and a

that the expected payoff for the defender is high



Table 7. Comparison of performance measures between existing method and proposed method

(a)

(b)

Figure. 6 Mixed strategies nash equilibrium for defender:

(a) probability of defender for single trojan type and (b)

trust-HUB circuits

positive payoff value for the defender. It is inferred

compared to that of attacker’s payoff, which clearly

guarantees that the defender chooses the best

stratergy against attacker for detecting the Trojan

type. Thus proposed methodology is validated by the

Figure. 7 Expected pay off for defender

mathematical game theory approach for detecting the

presence of malicious anamoly in the digital circuit.

5. Comparision with existing techniques

The existing machine learning technique are

classified into supervised and unsupervised learning.

The supervised learning algorithm like SVM based

classifier [11], Multilayer neural network [13]

method and the unsupervised learning algorithm like

boundary net structure [15], PL-HTD [16] are

compared with deep learning scheme. The proposed

deep learning technique are more focused, since it

overcomes the drawback of the machine learning

algorithm and it also improves the performance

metric. The proposed method is compared with the

related work for the Trust–HUB circuit and the

performance metric evaluated are summarized in

Table 7. It is observed that, comparing with MNN the

deep learning algorithm performs better with an

increase in average TPR by 5.69 % and average

accuracy by 52.27 %. As for PL- HTD technique, the

average value of TPR and accuracy of our method

increases by 49.50%, and 1.27% respectively. It is

deduced that the higher TPR leads to better detection

of Trojan nets and the rate of misjudgement of

Trust-HUB

circuits

True positive rate(TPR) True negative rate (TNR) Accuracy (%)

[13] [15] [16] Ours [13] [15] [16] Ours [13] [15] [16] Ours

RS232-T1000 100 100 50 93.3 24.00 98.2 96.43 82.9 32.58 98.4 94.85 95.23

RS232- T1100 78.00 69.4 45.45 100 25.00 96.8 96.43 91.57 30.36 93.8 94.50 96.08

RS232- T1200 91.00 100 46.15 98.00 55.00 95.8 97.14 89.17 58.79 96.3 94.56 95.12

RS232-T1300 86.00 100 57.14 96.5 65.00 99.7 96.04 72.08 66.93 99.7 95.06 95.05

RS232- T1400 100 100 41.67 98.00 15.00 97.1 96.40 85.93 27.07 97.5 94.14 96.82

RS232- T1500 82.00 97.4 45.45 94.05 47.00 97.5 96.45 92.76 51.24 97.5 94.54 97.63

RS232- T1600 100 96.4 44.44 97.00 28.00 98.3 96.11 88.24 34.23 98.1 94.52 95.63



genuine nets to be reduced, which highlights the

feature extraction in the extraction phase.

It is also inferred that deep learning based Trojan

detection performs better in TPR metric compared to

the existing machine learning techniques, which

indicated the nets affected by Trojan are perfectly

classified. Although the resulting TNR was to some

extend behind the boundary based and PL-HTD

methods, but the normal nets of the proposed method

focus on feature extraction for classifying the Trojan

net without compromising the classification of

genuine nets subsequently reduces the manual

computation. Compared with boundary based our

proposed method average TPR increases by 1.95 %

but the average accuracy is only 1.45 % smaller but

the overall accuracy of our scheme is 96.25% for

ISCAS and Trust –HUB circuit without

compromising the reliability of the design. To

summarize, it is analyzed that among the classifiers

(MNN, PL-HTD, Boundary based), the proposed

deep learning technique performs the best in terms of

the classification rate with the optimal feature set.

6. Conclusion

In this work a deep learning-based hardware

Trojan detection and classification technique is

proposed which develops algorithms for extracting

the circuit parameters. The structural reports of the

gate level netlist are generated in this work to extract

Trojan features and make the detection process a

static. The simulation is demonstrated on ISCAS’85,

ISCAS’89 and Trust-HUB circuits which shows that

the proposed technique achieves an average accuracy

of 96.25% and average True positive rate of 96.69%

at minimum processing time. The deep learning-

based detection technique is also validated using

game theoretical approach which provides maximum

probability for the defender by ensuring the

probability of detecting the Trojan in the benchmark

circuit. The major challenges of the algorithms

developed are addressing a specific threat model

attacks and providing the solutions. It has to be

extended and trained for different attack or threat

models. In future work would be incorporating the

delay parameters as a static timing analysis in the

proposed technique for detection process and also

modelling the denoising auto encoder architecture to

minimize the prediction error on a supervised

learning scheme.

Conflicts of Interest

The authors declare no conflict of interest.

Author Contributions

Conceptualization, methodology, software,

validation, writing, original draft preparation,

Priyatharishini Murugesan; review and editing,

supervision, project administration, Nirmala Devi

Manickam.

Acknowledgments

Currently the article is not receiving any funds

under any organization.

References

[1] S. Bhasin and F. Regazzoni, “A survey on

hardware Trojan detection techniques”, In: IEEE

International Symposium on Circuits and

Systems (ISCAS), pp. 2021-2024, 2015.

[2] M. Tehranipoor and F. Koushanfar, “A survey

of hardware Trojan taxonomy and detection”,

IEEE Design and Test of Computers, Vol. 27,

No. 1, pp. 10–25, 2010.

[3] H. Salmani, M. Tehranipoor and J. Plusquellic,

“A novel technique for improving hardware

Trojan detection and reducing Trojan activation

time”, IEEE Transactions on Very Large Scale

Integration (VLSI) Systems, Vol. 20, No. 1, pp.

112–125, 2011.

[4] M. Banga and M. Hsiao, “A region based

approach for the identification of hardware

Trojans”, IEEE International Symposium on

Hardware Oriented Security and Trust (HOST),

pp. 40–47, 2010.

[5] D. K. Karunagaran and N. Mohankumar,

“Malicious Hardware Trojan Detection by Gate

level minimization 90nm Technology”, In: Proc.

of Fifth International Conf. on Computing,

Communications and Networking Technologies

(ICCCNT), pp. 1-7, 2014.

[6] S. Dupuis, B. A. Papa Sidi, G. Di. Natale, M. L.

Flottes, and B. Rouzeyre, “A novel hardware

logic encryption technique for thwarting illegal

overproduction and hardware Trojans”, In: 2014

IEEE 20th International On-Line Testing

Symposium (IOLTS), pp. 49-54, 2014.

[7] B. Liu and B. Wang, “Embedded reconfigurable

logic for ASIC design obfuscation against

supply chain attacks”, In: Proc. of Design,

Automation & Test in Europe Conf. &

Exhibition (DATE), pp. 1-6, 2014.

[8] M. Priyatharishini, M. Nirmala Devi, “A

Compressive Sensing based optimal test pattern

generation for Hardware Trojan Detection”,

International Journal of Electrical and



Computer Engineering, Vol. 9, No.5, pp. 4035-

4043, 2019.

[9] J. Popat, U. Mehta, “Transition probabilistic

approach for detection and diagnosis hardware

Trojan in combinational circuits”, In: Proc. of

IEEE Annual India Conf. (INDICON), pp. 1-6,

2016.

[10] H. Salmani, “COTD: Reference-Free Hardware

Trojan Detection and Recovery Based on

Controllability and Observability in Gate-Level

Netlist”, IEEE Transactions on Information

Forensics and Security, Vol. 12, No. 2, pp. 338-

350, 2017.

[11] K. Hasegawa, M. Oya, M. Yanagisawa, and N.

Togawa, “Hardware Trojans Classification for

Gate-level Net-lists based on Machine

Learning”, In: IEEE 22nd International

Symposium on On-Line Testing and Robust

System Design (IOLTS), pp. 203-206, 2016.

[12] K. Hasegawa, M. Yanagisawa, and N. Togawa,

“Trojan-feature Extraction at Gate-level Net-

lists and Its Application to Hardware-Trojan

Detection Using Random Forest Classifier”, In:

IEEE International Symposium on Circuits and

Systems (ISCAS), pp. 1-4 ,2017.


“Hardware Trojans Classification for Gate-level

Net-lists Using Multi-Layer Neural Networks”,

In: IEEE 23rd International Symposium on On-

Line Testing and Robust System Design (IOLTS),

pp. 227-232 IEEE, 2017.

[14] O. I. Abiodun, A. Jantan, A. E. Omolara, K. V.

Dada, A. M. Umar, O. U. Linus, H. Arshad, A.

A. Kazaure, U. Gana, and M. U. Kiru,

“Comprehensive review of artificial neural

network applications to pattern recognition”,

IEEE Access, Vol. 4, No. 7, pp. 158820-158846,

2019.


“A hardware-Trojan classification method

utilizing boundary net structures”, In: Proc. of

IEEE International Conf. on Consumer

Electronics (ICCE), pp. 1-4, 2018.

[16] C. Dong, Y. Liu, J. Chen, X. Liu, W. Guo, and

Y. Chen, “An Unsupervised Detection

Approach for Hardware Trojans”, IEEE Access.

2020. (In press)

[17] I. Goodfellow, Y. Bengio, and A. Courville,

Deep learning, MIT press, Cambridge, 2016.

[18] Introduction to deep learning, Available:

http://introtodeeplearning.com.

[19] J. Graf, “Trust games: How game theory can

guide the development of hardware Trojan

detection methods”, In: IEEE International

Symposium on Hardware Oriented Security and

Trust (HOST), pp. 91-96, 2016.

[20] C. A. Kamhoua, H. Zhao, M. Rodriguez, and K.

A. Kwiat, “A Game-Theoretic Approach for

Testing for Hardware Trojan”, IEEE

Transactions on Multi-Scale Computing Systems,

Vol. 2, No.3, pp. 199-210, 2016.

[21] S, R. Hasan, C. A. Kamhoua, K. A. Kwiat, L.

Njilla, “A Novel Framework to Introduce

Hardware Trojan Monitors using Model

Checking Based Counterexamples: Inspired by

Game Theory”, In: IEEE 61st International

Midwest Symposium on Circuits and Systems

(MWSCAS), pp. 853-856, 2018.

[22] W. Saad, A. Sanjab, Y. Wang, C. A. Kamhoua,

K. A. Kwiat, “Hardware Trojan detection game:

A prospect-theoretic approach”, IEEE

Transactions on Vehicular Technology, Vol. 66,

No. 9, pp. 7697-7710, 2017.

[23] J. Graf, W. Batchelor, S. Harper, R. Marlow, E.

Carlisle, P. Athanas, “A practical application of

game theory to optimize selection of hardware

Trojan detection strategies”, Journal of

Hardware and Systems Security, Vol. 4, No. 2,

pp. 98-119, 2020.

An Optimal Feature Extraction using Deep Learning ...

Documents

Transcript of An Optimal Feature Extraction using Deep Learning ...